Helicopter Simulator Performance Prediction Using the Random Forest Method

INTRODUCTION: Different aspects of the aviation system, such as pilot's fitness, supervision, and working conditions, interact to produce or protect against flight safety hazards. Machine learning methods such as Random Forests may help identify system characteristics with the potential to affect flight safety from the large number of candidate predictors that results when multiple system levels are considered simultaneously. METHODS: There were 54 pilot-related and occupational candidate predictors of simulator flight performance in 2 malfunction scenarios completed by 51 male European helicopter emergency medical services pilots derived from pilots'self-report questionnaires and aeromedical examination records. In a cross-sectional explorative analysis, the Random Forest method was used to screen for informative predictors. Predictors scoring above the critical threshold for the conditional permutation variable importance (VI) statistic were selected. RESULTS: In five predictors, the VI statistic averaged across 2000 Random Forest runs exceeded the selection threshold: higher perceived rewards (VI = 0.0691) and predictability (VI = 0.0501) at work were associated with higher performance scores, and higher physiological dysregulation (VI = 0.0495) and alanine aminotransferase (VI = 0.0224) with lower scores. Performance also differed between the simulators at the two training sites (VI = 0.0298). DISCUSSION: Random Forests may usefully complement previously applied methods for the identification of human factors safety hazards. The identified performance predictors suggest further areas with potential for safety improvements.